Optimization of Machine Learning Pipelines with Singularity Containers on the HPC
Vogel, Bryan (Sensors and Algorithms Branch, Modeling & Simulation Division,
Research & Technology Integration Directorate, DEVCOM C5ISR Center)
Successes - Container Implementation and/or Deployment
Singularity containers provide a simple way to set up machine learning environments on the DoD HPC systems. This presentation investigates deploying and optimizing machine learning pipelines with Singularity containers. The implementation of input pipelines for efficiently reading and preprocessing image data to accelerate the training and inference processes will be discussed. Tensorflow 2, a machine learning library, has a useful data API which will be introduced, and the application of the API specifically on the DoD HPC resources will be demonstrated. Comparisons of runtime between the naïve approach, and the optimized approach will show an improvement in both the training and inference speeds. This enables faster iterations on tuning and testing new model architectures and reduces the amount of compute hours required for these processes.